Archive for infinite variance estimators

distributed evidence

Posted in Books, pictures, Statistics, University life with tags , , , , , , , , , , , , , , , , , , on December 16, 2021 by xi'an

Alexander Buchholz (who did his PhD at CREST with Nicolas Chopin), Daniel Ahfock, and my friend Sylvia Richardson published a great paper on the distributed computation of Bayesian evidence in Bayesian Analysis. The setting is one of distributed data from several sources with no communication between them, which relates to consensus Monte Carlo even though model choice has not been particularly studied from that perspective. The authors operate under the assumption of conditionally conjugate models, i.e., the existence of a data augmentation scheme into an exponential family so that conjugate priors can be used. For a division of the data into S blocks, the fundamental identity in the paper is

p(y) = \alpha^S \prod_{s=1}^S \tilde p(y_s) \int \prod_{s=1}^S \tilde p(\theta|y_s)\,\text d\theta

where α is the normalising constant of the sub-prior exp{log[p(θ)]/S} and the other terms are associated with this prior. Under the conditionally conjugate assumption, the integral can be approximated based on the latent variables. Most interestingly, the associated variance is directly connected with the variance of

p(z_{1:S}|y)\Big/\prod_{s=1}^S \tilde p(z_s|y_s)

under the joint:

“The variance of the ratio measures the quality of the product of the conditional sub-posterior as an importance sample proposal distribution.”

Assuming this variance is finite (which is likely). An approximate alternative is proposed, namely to replace the exact sub-posterior with a Normal distribution, as in consensus Monte Carlo, which should obviously require some consideration as to which parameterisation of the model produces the “most normal” (or the least abnormal!) posterior. And ensures a finite variance in the importance sampling approximation (as ensured by the strong bounds in Proposition 5). A problem shared by the bridgesampling package.

“…if the error that comes from MCMC sampling is relatively small and that the shard sizes are large enough so that the quality of the subposterior normal approximation is reasonable, our suggested approach will result in good approximations of the full data set marginal likelihood.”

The resulting approximation can also be handy in conjunction with reversible jump MCMC, in the sense that RJMCMC algorithms can be run in parallel on different chunks or shards of the entire dataset. Although the computing gain may be reduced by the need for separate approximations.

a common confusion between sample and population moments

Posted in Books, Kids, R, Statistics with tags , , , , , , , on April 29, 2021 by xi'an

sampling-importance-resampling is not equivalent to exact sampling [triste SIR]

Posted in Books, Kids, Statistics, University life with tags , , , , , , on December 16, 2019 by xi'an

Following an X validated question on the topic, I reassessed a previous impression I had that sampling-importance-resampling (SIR) is equivalent to direct sampling for a given sample size. (As suggested in the above fit between a N(2,½) target and a N(0,1) proposal.)  Indeed, when one produces a sample

x_1,\ldots,x_n \stackrel{\text{i.i.d.}}{\sim} g(x)

and resamples with replacement from this sample using the importance weights

f(x_1)g(x_1)^{-1},\ldots,f(x_n)g(x_n)^{-1}

the resulting sample

y_1,\ldots,y_n

is neither “i.” nor “i.d.” since the resampling step involves a self-normalisation of the weights and hence a global bias in the evaluation of expectations. In particular, if the importance function g is a poor choice for the target f, meaning that the exploration of the whole support is imperfect, if possible (when both supports are equal), a given sample may well fail to reproduce the properties of an iid example ,as shown in the graph below where a Normal density is used for g while f is a Student t⁵ density:

improved importance sampling via iterated moment matching

Posted in Statistics with tags , , , , on August 1, 2019 by xi'an

Topi Paananen, Juho Piironen, Paul-Christian Bürkner and Aki Vehtari have recently arXived a work on constructing an adapted importance (sampling) distribution. The beginning is more a review than a new contribution, covering the earlier work by Vehtari, Gelman  and Gabri (2017): estimating the Pareto rate for the importance weight distribution helps in assessing whether or not this distribution allows for a (necessary) second moment. In case it does not (seem to), the authors propose an affine transform of the importance distribution, using the earlier sample to match the first two moments of the distribution. Or of the targeted function. Adaptation that is controlled by the same Pareto rate technique, as in the above picture (from the paper). Predicting a natural objection as to the poor performances of the earlier samples, the paper suggests to use robust estimators of these moments, for instance via Pareto smoothing. It also suggests using multiple importance sampling as a way to regularise and robustify the estimates. While I buy the argument of fitting the target moments to achieve a better fit of the importance sampling, I remain unclear as to why an affine transform would change the (poor) tail behaviour of the importance sampler. Hence why it would apply in full generality. An alternative could consist in finding appropriate Box-Cox transforms, although the difficulty would certainly increase with the dimension.

Gibbs clashes with importance sampling

Posted in pictures, Statistics with tags , , , , , on April 11, 2019 by xi'an

In an X validated question, an interesting proposal was made: at each (component-wise) step of a Gibbs sampler, replace simulation from the exact full conditional with simulation from an alternate density and weight the resulting simulation with a term made of a product of (a) the previous weight (b) the ratio of the true conditional over the substitute for the new value and (c) the inverse ratio for the earlier value of the same component. Which does not work for several reasons:

  1. the reweighting is doomed by its very propagation in that it keeps multiplying ratios of expectation one, which means an almost sure chance of degenerating;
  2. the weights are computed for a previous value that has not been generated from the same proposal and is anyway already properly weighted;
  3. due to the change in dimension produced by Gibbs, the actual target is the full conditional, which involves an intractable normalising constant;
  4. there is no guarantee for the weights to have finite variance, esp. when the proposal has thinner tails than the target.

as can be readily checked by a quick simulation experiment. The funny thing is that a proper importance weight can be constructed when envisioning  the sequence of Gibbs steps as a Metropolis proposal (in the dimension of the target). Sad enough, the person asking the question seems to have lost interest in the issue, a rather common occurrence on X validated!

%d bloggers like this: